[Qemu-devel] Re: [PATCH] Always swap endianness in DBDMA

[Alexander Graf]

On 23.12.2009, at 11:06, Aurelien Jarno wrote:

> On Wed, Dec 23, 2009 at 10:29:53AM +0100, Alexander Graf wrote:
>> 
>> On 23.12.2009, at 07:12, Aurelien Jarno wrote:
>> 
>>> On Tue, Dec 22, 2009 at 02:45:17PM +0100, Alexander Graf wrote:
>>>> When we get an MMIO request, we always get variables in host endianness. 
>>>> The
>>>> only time we need to actually reverse byte order is when we read bytes from
>>>> guest memory.
>>>> 
>>>> Apparently the DBDMA implementation is different there. A lot of the logic
>>>> in there depends on values being big endian. Now, qemu does all the 
>>>> conversion
>>>> in the MMIO handlers for us already though, so it turns out that we're in
>>>> the same byte order from a C point of view, but cpu_to_be32 and be32_to_cpu
>>>> end up being nops.
>>>> 
>>>> This makes the code work differently on x86 (little endian) than on ppc 
>>>> (big
>>>> endian). On x86 it works, on ppc it doesn't.
>>>> 
>>>> This patch (while being seriously hacky and ugly) makes dbdma emulation 
>>>> work
>>>> on ppc hosts. I'll leave the real fixing to someone else.
>>> 
>>> I have to say I found it too hacky to be included in QEMU. I would
>>> prefer if someone can provide a real fix.
>> 
>> Looking at hw/pci_host_template.h I'm actually more confident now that this 
>> is actually the correct solution. We don't want big endian or little endian, 
>> we want reversed endian variables.
>> 
> 
> I guess it's the same problem, that is the bswap is there to emulate the
> fact that the bus is connected backward. Then I guess that the bswap 
> should also depends on TARGET_WORDS_BIGENDIAN, even if this file is 
> currently only compiled for big endian targets. Also a comment might be 
> a good idea.

Yeah, I was thinking of a helper function to convert that, so we don't have 
#ifdef's cluttered all over the code :-)

Alex